The agent could retrieve facts from last week, but the collaboration itself never improved. An analogy I like is that it’s a “smart stranger with a better notebook.”
The problem isn't retrieval. There are lots of tools that store and retrieve, to varying degrees of success. The problem is that nothing changes about how the agent works with you. It recalls what you said without learning how you think.
Solitaire is my attempt at the next layer. It's identity infrastructure for AI agents, and is designed to make the interaction itself improve over time.
What it does beyond memory: - Behavioral genome: disposition traits (observance, assertiveness, warmth, etc.) that evolve from real interaction, not static config. - Experiential memory: encodes how sessions felt, not just what was said. - Autonomous self-improvement: retrieval weights adjust based on what proved useful. The knowledge graph self-heals (contradiction detection, confidence rescoring, entity relinking). - Anticipatory retrieval: predicts what context you'll need and preloads it before you ask. - Guided onboarding: new users build a partner through conversation, not a JSON file. - Memory-compatibility: use Solitaire’s built-in memory or bring your own existing memory tools. - Model-compatibility: many work out of the box, we’re building towards integrating others. - Your data stays local, so you have full data sovereignty.
600+ sessions, 14,000+ accumulated entries in real production use and counting. I have two research papers coming out of the longitudinal work.
Repo: https://github.com/PRDicta/Solitaire-for-Agents
I know the agent memory space is popular, but memory is one component. Identity is the thing nobody's building. The agent should get better at working with you, not just better at remembering.
Would especially welcome feedback on: 1. Does the identity vs. memory distinction make sense to you? 2. Where do current memory tools break down for you in practice? 3. What would make something like this worth integrating into your agent stack? 4. What am I missing?
dictadev•1h ago
Anyways:
--> The comment (from dishitarocks):
"The "identity layer" framing is interesting. Most memory tools just do store and retrieve. Considering that the space is also getting crowded, how are you thinking about identity differently? Is it more like persistent preferences and personality that carry across sessions, or is it tracking what the agent has done and learned over time?
Curious how this handles conflicts too. If an agent's stored identity says one thing but the current conversation context says another, which wins?"
--> My answer:
Solid question, which I was pretty sure I'd see.
Memory tools at present record things, but don't learn from anything.
They're a strong record, but the mind behind it stays the same. And 50 sessions deep feels the same as five.
Identity, or what I'm trying to build, is about growth, and compounding.
The best way to view it is true continuity: where the disposition traits I built (warmth, humor, and a bunch more) will dynamically shift over time as you work with it. It's natural, and it's automatic. You can manually tune anything you'd like, but I've found that finding that rhythm is more genuine, somehow. To your point about learned work efforts, that's also part of that compounding. "I need another X, this time on client Y" = "Got it!" That's been my experience, and it's a smooth one.
I'm a huge fan of memory tools, and this project even comes with one (should you need it), but they're really on the first step towards something greater.
On conflicts: current convo wins, always. User input takes precedence, because you absolutely *know* that you like peanut butter and chocolate, whatever the record says about you hating that combination.
But the system doesn't just silently discard that contradiction. It flags it, rescores confidence, and keeps both with temporal context/tagging so you can trace how that thinking evolved.
Maybe you said it sarcastically once, and it was filed as a fact? Whatever the genesis, there is a path to consolidation and correction (and a way to fully delete anything causing real problems).